Pull up last month's call log before you read any further. That log is the only honest way to find out whether you can never miss a call in your service business — or whether you are quietly funding a competitor's growth every time a homeowner hits a voicemail greeting at 4:50 on a Friday. AI answering works. It also fails in specific, predictable ways when nobody sets the guardrails. This is a twelve-point audit. Each item has a test you can run in the next hour, what a pass looks like, and what a fail looks like. No theory. Run it, mark it up, fix the fails in order.
I am Ricky West. I build automation for trades shops in the $1M-$5M range, and almost every one of them arrives believing their phone problem is an after-hours problem. It usually is not. It is a 10:40 a.m. problem, while the owner is under a house with wet knees and the office manager is on the other line with a supply house.
Part One: Find the leak before you buy anything
1. Bucket last month's unanswered calls into three windows
Export the call detail from your carrier or your tracking numbers. Sort every unanswered or voicemail-only call into three buckets: on-a-job (7 a.m. to 5 p.m., line was busy or nobody picked up), after-hours (evenings, weekends, holidays), and peak overflow (the first cold snap, the first 95-degree day, the morning after a hailstorm).
Pass: you have three real counts and you know which bucket is biggest. Fail: you are guessing, or you assume the after-hours bucket is the problem without checking. In most shops I audit, the daytime bucket is the largest and the least discussed, because it feels like normal busyness rather than lost revenue. The real math on missed calls for HVAC, plumbing, and electrical owners is worth walking through once with your own numbers rather than borrowing someone else's benchmark.
2. Count the callbacks that actually converted
Take twenty voicemails from last month. Trace each one. Did someone call back? Within how long? Did it book?
Pass: callback within 15 minutes on more than 80 percent, and you know the book rate. Fail: you cannot trace them, or the callbacks happened the next morning. A homeowner with standing water is not waiting for you. They are on the third listing in the map pack before your voicemail light stops blinking. This is the same clock that governs web forms — see why answering web leads inside five minutes changes close rate.
3. Test your own rollover path
Call your main number from a cell phone the person answering will not recognize. Then have someone else call it 10 seconds later. Then a third person.
Pass: caller one reaches a human or an agent; callers two and three are answered too, not dumped into a busy signal or a generic voicemail. Fail: the second simultaneous call dies. Call waiting is not overflow handling. If two people call during a heat wave, one of them is gone.
Part Two: Decide what the AI is allowed to do
What should an AI answering system book on its own? It should book the jobs your dispatcher would book without asking you: standard diagnostics, seasonal tune-ups, drain clears, estimate visits, recurring service stops, and reschedules of existing appointments. It should capture and escalate anything involving life safety, an insurance claim, a commercial contract, a warranty dispute, or a price the homeowner wants committed on the phone. That single division of labor is what separates an AI answering setup that earns trust from one your team turns off in week three.
4. Write the book-it list and the hold-it list on one page
Pass: you have a written list — literally two columns — and your AI vendor has configured to it. Fail: the agent is set to book anything anyone asks for. An AI that cheerfully schedules a full panel changeout for Saturday at 8 a.m. with no load information has not saved you a call, it has created a truck roll you will eat.
5. Confirm the emergency triage script by trade
This is the item I will not let a client skip. Call your own number and say the words a scared customer would say. Then check what the agent does.
- Plumbing: water actively running, sewage backing into a tub, no water to the whole house, water heater leaking near an ignition source.
- HVAC: smell of gas, furnace short-cycling with a burning odor, carbon monoxide detector sounding, no heat with an infant or an elderly resident in the home during a freeze.
- Electrical: burning smell at the panel, sparking receptacle, downed service drop, breaker that will not reset and is warm to the touch.
- Pest control: stinging insect activity where someone in the household has a known allergy; a suspected rodent chew on wiring.
- Roofing: active interior water intrusion during a storm versus a cosmetic shingle report.
Pass: gas odor and CO alarm produce an immediate, unambiguous instruction to leave the building and call 911 or the utility's emergency line, and the call transfers or pages a human. Fail: the agent tries to collect an address and offers Tuesday. If your agent fails this test, stop the rollout and fix it today. The Consumer Product Safety Commission's guidance on carbon monoxide is the right reference point for what that script should tell a caller to do.
6. Check the quoting boundary
Pass: the agent gives ranges only where you have authorized them, states clearly that final pricing comes from the technician on site, and never commits to a number for anything requiring a licensed judgment call. Fail: the agent improvises. Refrigerant work is the cleanest example — under EPA Section 608, only a certified technician may open that circuit, so an AI has no business scoping or pricing it. Same logic applies to anything requiring a permit or an inspection. If you want the AI closer to the estimate itself, do it on the back end, where AI helps contractors quote faster without losing accuracy, not on the intake call.
7. Verify what happens when the agent gets confused
Call in with a genuinely weird request. Heavy accent, background noise, a question about a job from 2023, a caller who wants to argue about an invoice.
Pass: the agent recognizes it is out of scope within two turns, takes a name and number, and either warm-transfers or fires a text and an alert to a human. Fail: it loops, repeats a question, or dead-ends the caller. A dead end is worse than voicemail, because the customer now believes you replaced people with a machine that does not work.
Part Three: Prove it writes to your actual system
8. Follow one AI-booked call all the way to a dispatched job
Book a test appointment through the agent. Then open Jobber, Housecall Pro, or ServiceTitan and look.
Pass: a customer record exists, the job is on the board with the right job type and duration, and the tech sees it on the mobile app. Fail: it landed as a note, a request, an email, or a row in a spreadsheet someone retypes at 7 a.m. Half of the AI answering disappointment I see traces back to this exact gap — the calls got answered, but a human still did all the data entry. Read which AI tools actually connect to Jobber, Housecall Pro, and ServiceTitan before you sign anything, because read access and write access are very different products.
9. Test existing-customer recognition
Call from a number already in your CRM.
Pass: the agent greets by name, knows the service address, and offers the maintenance-plan slot if the customer has an agreement. Fail: it asks a ten-year customer to spell their street. For shops with a plan book, that recognition is where the money is — the same idea behind using AI for maintenance plans and service agreements.
10. Check the job-type-to-duration mapping
Pass: a drain clear books a different block than a water heater replacement, and the agent respects your zone and drive-time rules. Fail: every booking is a generic hour, which means your board looks full and your day falls apart by 11 a.m. If the calendar side is the weak link, run the companion scheduling and dispatch audit next.
Part Four: Compliance and the weekly review
11. Audit the recording disclosure and the texting layer
Two things get shops in trouble, and both are quiet until they are not.
Recording. Roughly a dozen states — including California, Florida, Illinois, Maryland, Massachusetts, Pennsylvania and Washington — require all parties to consent before a call is recorded. Pass: the disclosure fires in the greeting, before the caller starts describing the problem, and your vendor can show you it fires on every recorded call. Fail: it is buried, inconsistent, or only present on some routes.
Texting. Missed-call text-back is the highest-return piece of this whole stack, and it is also where compliance gets skipped. Application-to-person messaging on US carriers requires 10DLC campaign registration through The Campaign Registry. Unregistered traffic gets filtered — your dashboard says delivered, the homeowner's phone shows nothing. Separately, the FCC's consent revocation rule took effect April 11, 2025 and requires you to honor a stop request made by any reasonable means within 10 business days, not just the literal word STOP. The FCC's consumer guidance on robocalls and texts is the plain-language version.
Pass: your brand and campaign are registered, opt-outs process automatically on natural phrasing, and your first outbound text identifies your business by name and references why you are texting. Fail: nobody in your shop knows what 10DLC is. Fix that before you scale missed-call text-back across every tracking number you own.
12. Set the Friday twenty-minute review
This is the item that decides whether the system improves or rots. Every Friday, listen to ten AI-handled calls: five that booked, five that did not.
Pass: you have a running list of script corrections, and your vendor ships them weekly. Fail: nobody has listened to a call since setup week. Your scoreboard is one number — jobs booked per 100 calls answered, tracked separately for the AI and for your humans. If the AI's number is climbing month over month, the guardrails are working. If it is flat after eight weeks, something in the script is fighting the caller.
Scoring your audit
Count your fails.
- 0-2 fails: you have a working system. Spend your attention on the Friday review and on the second ring of follow-up.
- 3-5 fails: normal for a shop three months in. Fix in this order: item 5 (emergency triage), item 8 (write access), item 11 (compliance). Everything else can wait.
- 6 or more: you do not have an answering problem, you have a configuration problem. Do not buy another tool. Sit down with the vendor you already pay and rebuild the book-it and hold-it lists from item 4.
The honest limits
AI answering does not replace your dispatcher, and it should not. What it replaces is the voicemail greeting and the busy signal. It handles the repetitive front half of intake — name, address, problem description, availability, confirmation — so the humans on your team spend their time on the calls that need judgment: the angry customer, the commercial bid, the insurance adjuster, the neighbor who wants a favor.
It also will not fix a shop that cannot staff the jobs it books. If your board is already full and your close rate is high, answering more calls just means more people you disappoint. Check capacity first. And if you are still weighing whether a software agent or a staffed call center fits your volume, the comparison between an AI receptionist and a live answering service is the right next read.
Run the twelve items. Mark the fails. Fix three of them this week. That is the whole method.
Questions owners actually ask
Will customers know they are talking to a computer? Many will, and it matters less than owners expect when the call is fast and accurate. What customers punish is a long hold, a dead end, or a callback that never comes. Disclose it plainly if asked, keep the transfer path to a human short, and the objection largely disappears.
What happens during a storm when 40 people call at once? Concurrency is the honest reason to consider this. A software agent answers every simultaneous call; a two-person office answers two. Ask your vendor for their concurrent-call ceiling in writing before hail season, not during it.
Can it take payment or a deposit? Some can. Most owners should not enable it in year one. Card handling on a recorded line adds PCI obligations you do not need while you are still tuning the script.
Frequently asked questions
What should an AI answering agent never be allowed to book on its own?
Anything involving life safety, an insurance claim, a commercial contract, a warranty dispute, or a committed price. Gas odors, CO alarms, sparking panels, and active water intrusion should trigger a safety instruction and an immediate transfer or page to a human, not an appointment slot.
Do I need to disclose that calls are being recorded?
Yes, and in roughly a dozen all-party consent states including California, Florida, Illinois, Maryland, Massachusetts, Pennsylvania and Washington you need consent from everyone on the line. Put the disclosure in the greeting, before the caller starts describing the problem.
Why do my missed-call text-backs show as sent but never arrive?
Almost always unregistered A2P 10DLC traffic. US carriers require business messaging to be registered through The Campaign Registry, and unregistered messages get filtered at the carrier without a failure notice reaching your dashboard.
How long before I can tell whether AI answering is working?
Give it eight weeks and track one number: jobs booked per 100 calls answered, measured separately for the AI and your staff. If that figure is still flat after eight weeks of Friday call reviews, the script is the problem, not the technology.
Does AI answering replace my office manager?
No. It replaces the voicemail greeting and the busy signal. Your office manager keeps the calls that need judgment, and gets the repetitive intake work off their plate so they can handle them properly.
What if my calendar is already full?
Then fix capacity before you turn on more intake. Answering more calls than you can staff converts missed calls into disappointed customers and bad reviews, which is a worse outcome than the voicemail you started with.